Rapid and environmentally friendly wine analysis using Vis/NIR Spectroscopy and Support Vector Machine regression

Visible and Near Infrared (Vis/NIR) spectroscopy has been investigated extensively during the last decade to measure quality parameters in wine. This technique has the advantages of being fast, non-destructive, requires no sample preparation and is chemical reagent-free, leading to no generation of residues. However, calibration models developed for the determination of some quality parameters in wine can exhibits non-linearities, leading to poor predictive performances when linear regression techniques are used. Therefore, a regression method that is able to work with non-linearities is an interesting alternative to improve the robustness and the predictive performance of the calibration models. The aim of this study was to calibrate and validate models that can be used to determine quality parameters in red wine using Vis/NIR spectroscopy and the Least Squares Support Vector Machine (LS-SVM) regression. A total of 72 experimental ‘Syrah’ red wines produced in the São Francisco River Valley, Brazil, were used in this study. Wine samples were kept at 33ºC and subjected to transflectance analysis using a benchtop spectrometer (XDS, FOSS NIRSystems, MD, USA) in the spectral range from 400 to 2500 nm. Four separate measurements were carried out on each sample and the four spectra were averaged to provide a mean spectrum. Standard Normal Variate (SNV) was used as spectral preprocessing and the samples were divided into calibration and validation sets using the SPXY (Sample set Partitioning based on joint x–y distances) algorithm. Multivariate calibration models were developed using LS-SVM regression. Predictive performance of the models was evaluated using the coefficient of determination in the external validation step (R2v) and the root mean square error of prediction (RMSEP). Reference analyses were accomplished for the determination of pH, total acidity (TA), volatile acidity (VA), total monomeric anthocyanins (TMA), total phenolic compounds (TPC), color intensity (CI), tonality (TO), total polyphenolic index (TPI), dry extract (DE) and alcohol content (AC). The predictive performance of the models are shown as follows in terms of R2v and RMSEP, respectively. pH: 0.80 and 0.08 (range: 3.25-3.86); TA: 0.73 and 7.68 mEq L-1 (range: 65.96-142.92 mEq L-1); VA: 0.71 and 0.82 mEq L-1 (range: 3.87-14.33 mEq L-1); TMA: 0.67 and 20.5 mg L-1 (range: 77.2-265.9 mg L-1); TPC: 0.72 and 212 mg L-1 (range: 1,233-2,835 mg L-1); CI: 0.95 and 0.78 (range: 7.07-21.97); TO: 0.87 and 0.03 (range: 0.26-0.79); TPI: 0.90 and 3.19 (range: 38.9-87.0); DE: 0.67 and 3.33 g L-1 (range: 15.50-43.85 g L-1); and AC: 0.99 and 0.22% (range: 5.92-17.81% v/v). The models obtained were able to predict with acceptable accuracy the wine quality parameters evaluated in this study. The results show that the Vis/NIR spectroscopy associated with LS-SVM has potential to provide rapid and accurate wine analysis without generation of dangerous residues.